This repo hosts a C++ and python implementation of the YOLOv9 state of the art object detection model, leveraging the TensorRT API for efficient real-time inference.

🚀 Usage

Python

cd <this project path>/python
python yolov9_trt.py --engine yolov9-c.engine --data images --outdir output

C++

cd <this project path>/build/release

# infer an image
yolov9-tensorrt.exe yolov9-c.engine test.jpg
# infer a folder(images)
yolov9-tensorrt.exe yolov9-c.engine data
# infer a video
yolov9-tensorrt.exe yolov9-c.engine test.mp4 # the video path

🛠️ Build

Python

The following command will install tensorrt for python:

cd <tensorrt installation path>/python
pip install cuda-python
pip install tensorrt-8.6.0-cp310-none-win_amd64.whl
pip install opencv-python

C++

Refer to our docs/INSTALL.md for detailed installation instructions.

• Cuda preprocessing: main branch
• Cpu preprocessing: cpu-preprocessing branch

🖥️ Requirement

• TensorRT
• CUDA, CudaNN
• C++ compiler with C++17 or higher support
• CMake 3.14 or higher
• OpenCV

🌱 Examples

Parameters for inference:

Confidence threshold : 0.2
NMS threshold : 0.3
Model : yolov9-e (with FP32)

Inference result:

Original image: https://www.flickr.com/photos/nicolelee/19041780

👏 Acknowledgement

This project is based on the following awesome projects:

• YOLOv9 - YOLOv9: Learning What You Want to Learn Using Programmable Gradient Information.
• TensorRT - TensorRT samples and api documentation.
• TensorRTx - Implementation of popular deep learning networks with TensorRT network definition API.

🔗 See also

• Yolov9-Bytetrack - Integration of YOLOv9 with ByteTracker using the TensorRT API.